AI BASED EEG WORK LOAD ASSESSMENT SYSTEM

MINOR PROJECT REPORT

Submitted by

HEMANTH MAMILLAPALLI(18UEEC0248)

DINESH TEJA LINGALA (18UEEC0232)

CHALAPATHI NAIDU MADDINENI(18UEEC0237)

Under the Guidance of

R.RAMASAMY

in partial fulfillment for the award of the degree

of

BACHELOR OF TECHNOLOGY

in

ELECTRONICS AND COMMUNICATION ENGINEERING

DECEMBER 2021
 BONAFIDE CERTIFICATE

Certified that this minor project report entitled ” AI BASED EEG WORK LOAD ASSESS-
MENT SYSTEM ” is the bonafide work of “ HEMANTH MAMILLAPALLI (18UEEC0248)
, DINESH TEJA LINGALA (18UEEC0232) and CHALAPATHI NAIDU MADDINENI
(18UEEC0237)” Who carried out this minor project work under my supervision.

SUPERVISOR HEAD OF THE DEPARTMENT

Mr.R.RAMASAMY Dr. P. ESTHER RANI

Assistant Professor Professor
Department of ECE Department of ECE

Submitted for Evaluation of Minor Project viva voce held on:−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

INTERNAL EXAMINER EXTERNAL EXAMINER

ii
 ACKNOWLEDGEMENT

We express our deepest gratitude to our respected Founder President and Chancellor Col.
Prof. Dr. R. Rangarajan, Foundress President Dr. Sagunthala Rangarajan, Chairperson
Managing Trustee and Vice President.

We are very thankful to our beloved Vice Chancellor Prof. Dr. S. Salivahanan for providing us
with an environment to complete the project work successfully.

We obligate our beloved Registrar Dr. E. Kannan for providing immense support in all endeavours.
I am thankful to our esteemed Dean Academics Dr. A. T. Ravichandran for providing a wonderful
environment to complete our work successfully.

We are extremely thankful and express the gratitude to our Dean Dr. V. Jayasankar for his valuable
guidance and support on completion of this project.

It is a great pleasure for us to acknowledge the assistance and contributions of our Head of the De-
partment Dr. P. Esther Rani, Professor for her useful suggestions, which helped us in completing
the work in time and we thank her for being instrumental in the completion of final year (7th sem)
with her encouragement and unwavering support during the entire course.

We are extremely thankful and pay the gratitude to our Minor Project Supervisor Mr.R.Ramasamy,
Assistant Professor for his valuable guidance and support on completing this minor project in pleasant
form.
We thank our department faculty, supporting staffs, family and friends for encouraging and supporting
throughout the study to complete this project.

HEMANTH MAMILLAPALLI
DINESH TEJA LINGALA
CHALAPATHI NAIDU MADDINENI

iii
 TABLE OF CONTENTS

ABSTRACT v

LIST OF ABBREVIATIONS vi

LIST OF FIGURES vii

1 INTRODUCTION 1
1.1 Introduction to EEG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 EEG signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 LITERATURE SURVEY 6

3 METHODOLOGY 9
3.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.1 Data Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Line noise filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Artifact Subspace Reconstruction (ASR) . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.5 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.5.1 Data labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.5.2 Training model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.6 Workload Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.6.1 Workload detection/recognition . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.6.2 System performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 CODE AND SIMULATION 15

5 RESULT AND DISCUSSION 21

6 CONCLUSION 29

7 FUTURE ENHANCEMENTS 30

REFERENCES 31

iv
 ABSTRACT

Individual’s mental and physical health’s as well as their performance, are affected by exces-
sive mental workload. The mental effort of operators doing vital activities must be monitored. EEG
can collect electrical signals produced by neural structures in the brain and provide information about
an operator’s mental state. It can be deduced the power distribution of these transmissions in various
frequency bands has changed. The mental workload has been assessed using this method. Because
of the poor signal to noise ratio, these noisy signals necessitate extensive filtering and pre-processing.
The accuracy of EEG-based workload assessment is influenced by factors like on an open-source work-
load dataset, STEW dataset, the performance of a workload assessment pipeline in relation to these
parameters in this negotiation. Furthermore, data gathered from individual lobes are used to analyse
workload assessment performance. Instead of utilising the entire brain, multiple frequency bands were
used instead the whole range of frequencies finally, for three-level workload classification, we compare
the performance of a number of classifiers. We approved our calculations to ensemble ANN calcu-
lation with neighbour-hood data given by EEG information base gathered during the execution of
Human errands. The Deep Learning calculation had given greatest precision contrasted with Machine
Learning Classifiers. By executing our calculation we got the greatest precision of 78 to 80 percent.
From this we can say that ANN Algorithm will be best suitable for calculating human brain activities.

v
 LIST OF ABBREVIATIONS

CV Computer Vision
EEG Electroencephalogram
AN N Artificial Neural Network
ASR Articraft subspace reconstruction
ST EW Simultaneous task EEG workload dataset
CN N Convolutional Neural Networks
BCI Brain Computer Interface
SN R Signal-to-Noise Ratio
AI Artificial Intelligence
ML Machine Learning

vi
 LIST OF FIGURES

1.1 Human Brain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Channel locations in Brain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Emotiv EPOC+ 14-channel EEG device . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3.1 Workload measurement pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3.2 EEG data recording pipelinet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Data set-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Data set-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

5.1 Code snippet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.2 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3 Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.4 Model accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.5 Model loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.6 Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.7 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.8 Imbalance Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

vii
 CHAPTER 1

INTRODUCTION

In the last decade, the usage of wearable devices for tracking our everyday physiological
activity, as well as their general availability, has exploded. Smartwatches/gears, as well as health and
fitness bands, are very popular these days . These devices are becoming increasingly feature-rich,
with inbuilt sensors for counting Steps, heart rate/pulse rate, and so on. One element of the human
body that may be tracked using a device is the heart. The human brain activity is a wearable device.
Electroencephalography (EEG) is a type of electroencephalography. the methods used to monitor
and record itAn electroencephalogram (EEG) is a recording of the brain’s spontaneous and rhythmic
electrical activity.The electrical potential created by the activation of numerous neurons in the human
brain is measured by these instruments. The individual who comes up with the idea first EEG was
used by a German physiologist to record human brain processes. In 1924, Hans Berger. Later, EEG
was used to assess patients in a clinical setting. and to diagnose disorders such as seizure, epilepsy,
and stress in the human brain as well as a variety of mental illnesses. In the previous decade, EEG was
restricted to the clinical setting, and EEG devices were both expensive and non-portable, despite the
fact that qualified and skilled EEG technicians were available. The acquisition of EEG necessitated
the use of individuals. However, with the advent of the new millennium and the EEG acquisition
devices are portable, often wireless, and simple to operate, with no need for professional assistance.
Electrodes are used to gather electrical signals from the brain in these devices.
The information concerning brain activities is contained in the signals recorded from the
brain.According to[3–5], relevant information in signals is found in These bands’ frequency ranges are
(0 – 4 Hz), (4 – 8 Hz), (8 – 13 Hz), and (13 – 30 Hz)and (more than 30 Hz), respectively. Fatigue,
workload, seizures, epilepsy, and stress are all possibilities. The band powers ratios of these four bands
can be evaluated With the advancement of technology, the usage of EEG has become more prevalent.
EEG technology has played a significant role because of its simplicity, ease of use, and high
time resolution of signals. In clinical practise, EEG is utilised as an indicator for the detection and
monitoring of conditions like epilepsy and sleep problems. EEG is a brain imaging technique that
employs electrodes attached to the scalp’s surface to identify and record electrical activity signals of
neuronal clusters in the cerebral cortex using precise electronic measurement technology, allowing re-

1
searchers to obtain information about the brain’s idea and cognition. Electrophysiological information
from the brain that is related to thinking and decision-making

Figure 1.1: Human Brain

The EEG is used in the research and development of rehabilitation equipment, such as
the development of brain-computer interface (BCI) and neuro feedback technologies to achieve the
recovery of patients’ motor cognition and other functions, and machine learning algorithms are often
used to decode EEG signals to accurately identify physiological or pathological conditions. However,
due to the lower spatial resolution and signal to noise ratio (SNR) of EEG signals, machine learning
decoding accuracy is limited, posing several difficulties in practical applications.

Figure 1.2: Channel locations in Brain

2
 Researchers have increasingly used new and efficient machine learning algorithms to EEG
decoding in recent years, and originally established its advantages over standard machine learning.
This covers the essential ideas of the deep learning algorithms currently used in EEG decoding, and
then introduces them. However, because the received signal is noisy, extracting meaningful information
is difficult. A lot of preprocessing is necessary. EEG isn’t the only way to record brain activity;
functional Magnetic Resonance Imaging (fMRI), functional Near Infrared Spectroscopy (fNIRS), and
Magnetoencephalography (MEG) are some of the other options for brain imaging [4]. All of them
differ in terms of characteristics, spatial and temporal resolution, portability, and cost. The majority
of fMRI machines are employed in medical settings since they are very expensive, not portable,
and work in very controlled surroundings.The method is tested in numerous typical EEG decoding
application settings, and the challenges encountered in EEG decoding analysis are discussed, as well
as the program’s future development.EEG is employed in professional settings as well as for personal
enjoyment. Gaming controls, mental problem-solving abilities testing, and a variety of other brain-
related applications are frequently provided by EEG device manufacturers. EEG has been used to
evaluate mental stress [5, 6], mental exhaustion [7], stress levels at construction sites [8], and workload
[9] in a number of investigations. We investigate the use of EEG to evaluate mental workload in this
thesis. We’re looking at which brain regions and EEG bands contribute the most to determining
mental workload. We look at how existing workload measurements perform when the number of
channels and EEG bands are reduced.

1.1 Introduction to EEG

One of the procedures used to obtain objective measurements of brain activity is electroen-
cephalography (EEG). It’s a type of neuroimaging technology that records electrical activity in the
brain. During different tasks, varying quantities of electrical activities occur in different regions of the
brain, such as the frontal lobe, parietal lobe, motor cortex, temporal lobe, and occipital lobe. Some
electrodes must be placed on these lobes in order to record these activity. These electrodes function
as recording channels for these devices. Channels receive data with very low amplitude, commonly
measured in microvolts (V). There are three different kinds of EEG. First, there’s the invasive brain
computer interface (BCI), which necessitates the use of electrodes These lobes require some atten-
tion.It has a very high temporal resolution. It depicts the dynamics of cognitive operations across the
course of the time period in which they occur. It may record in the range of hundreds to thousands
of milliseconds. It directly measures brain activities (voltage changes), and the data collected with
EEG devices is multidimensional. It appears to be two-dimensional data on the surface. The first is
temporal, which is defined by the change in voltage per unit time, or sampling rate, of an EEG device,
and the second is spatial, which is defined by the space covered by the device’s placed electrodes.
However, it is actually four-dimensional data, consisting of time, space, frequency, and power. These
electrodes are connected to these devices as channels for recording. The amplitude of data received
by channels is commonly measured in microvolts (V). EEG comes in three varieties. First, there’s

3
the invasive brain computer interface (BCI), in which electrodes are implanted beneath the skull in
the cerebral lobes. Second, a partial invasive brain computer interface in which electrodes are in-
serted inside the scalp but only on the surface of the brain lobes, rather than inside the brain. Third,
non-invasive brain computer interface, in which electrodes are placed on the scalp. In this thesis, We
focus on non-invasive brain computer interface technique as they are most common and safe for use
in application other than medicine.

Figure 1.3: Emotiv EPOC+ 14-channel EEG device

EEG is a well-known neuro imaging technology that is simple to record thanks to modern
devices’ portability and tolerance for patient mobility during a recording session. It has a very high
temporal resolution. It depicts the dynamics of cognitive operations throughout the course of a given
time period. It may record in the range of hundreds to thousands of milliseconds. It directly measures
brain activities (voltage changes), and the data collected with EEG devices is multidimensional. It
appears to be two-dimensional data on the surface. The first is temporal, which is defined by the
change in voltage per unit time, or sampling rate, of an EEG device, and the second is spatial,
which is defined by the space covered by the device’s placed electrodes. However, it is actually four-
dimensional data, consisting of time, space, frequency, and power . Many study questions, for example,
are unsuitable for EEG because of the great relevance of spatial resolution and the necessity to detect
sluggish cognitive processes. Low temporal and high spatial resolution are better suited to fMRI and
fNIRS.

1.2 EEG signals

When neurons activate, electrical pulses are generated in the nervous system, and potential
fluctuations occur at the scalp. EEG signal refers to the fluctuations in potential that can be monitored
with an EEG instrument. In relation to the event trigger, these measured signals correlate to distinct
lobes of the brain. Microvolts (V) are used to measure EEG signals. Electrodes implanted on the
head can be used to record these microvolts. The difference in electrical activity between the electrode

4
and the reference electrode is measured in microvolts. as well as the reference electrode Various
commercially available equipment, such as the NeuroSky MindWave Mobile 2 headset (single channel),
Muse 2 Headset (4-Channel), Emotiv EPOC+ (14-Channel), are utilised to record the electrical
potentials. All of these devices contain electrode channels that must be put on the scalp in accordance
with the 10-20 international electrode placement standard.

5
 CHAPTER 2

LITERATURE SURVEY

The following are various literature studies on Industrial Plant Monitoring, Automation, and
Control, which provide a brief overview of the research gaps, problem formulation, and aims..

Samarth Tripathi, Shrinivas Acharya, Ranti Dev Sharma Using Deep and Convolutional
Neural Networks for Accurate Emotion Classification on DEAP Dataset 2017 Emotion recognition is
an important field of research in Brain Computer Interactions. As technology and the understand-
ing of emotions are advancing, there are growing opportunities for automatic emotion recognition
systems. Neural networks are a family of statistical learning models inspired by biological neural
networks and are used to estimate functions that can depend on a large number of inputs that are
generally unknown. Research in the field of Emotion recognition and explore new techniques of using
the effectiveness of Neural Networks to classify user emotions using EEG signals from the DEAP
dataset

Mei and X. Xu EEG-based emotion classification using convolutional neural network 2017
Emotion classification based on brain–computer interface (BCI) systems is an appealing research
topic. Recently, deep learning has been employed for the emotion classifications of BCI systems. A
new deep network is proposed to classify EEG signals for emotion recognition. When embedding an
SAE structure into a CNN.

Hong Zeng , Chen Yang , Guojun Dai , Feiwei Qin EEG classification of driver mental states
by deep learning 2018 Driver fatigue is attracting more and more attention, as it is the main cause
of traffic accidents, which bring great harm to society and families. This paper proposes to use deep
convolutional neural networks, and deep residual learning, to predict the mental states of drivers from
electroencephalography (EEG) signals. Accordingly we have developed two mental state classification
models called EEG-Conv and EEG-Conv-R. A 5-layer convolution neural network is built to classify
the mental states of drive fatigue, and both classifiers are tested by raw EEG data. The classification

6
performances of these two models are compared to the classical SVM classifier and LSTM deep learn-
ing model with the same EEG.

S. Chambon A DEEP LEARNING ARCHITECTURE TO DETECT EVENTS IN EEG

SIGNALS DURING SLEEP 2018 Electroencephalography (EEG) during sleep is used by medical
professionals to evaluate various neurological disorders. We propose a deep learning method that pre-
dicts locations, durations and types of events in EEG time series. It relies on a convolutional neural
network that builds a feature representation from raw EEG signals. The proposed approach learns by
backpropagation to build a feature representation from the raw input signal(s) and to predict both
locations, durations andtypes of events. Numerical experiments on spindles and Kcomplexes detection
demonstrate that the proposed approachoutperforms previously published detection methods.

V. J. Lawhern, A. J. Solon, N. R.Waytowich, S. M Gordon, C. P. Hung EEG Net: A compact

convolutional neural network for EEG based brain–computer interfaces, Brain computer interfaces
2018(BCI) enable direct communication with a computer, using neural activity as the control signal.
EEGNet is a compact convolutional neural network for EEG-based BCIs. It is robust enough to learn
a wide variety of interpretable features over a range of BCI tasks. A Brain-Computer Interface (BCI)
enables direct communication with a machine via brain signals. Traditionally, BCIs have been used
for medical applications such as neural control of prosthetic limbs. Recent research has opened up the
possibility for novel BCIs focused on enhancing performance of healthy users.

J Thomas, L. Comoretto, J. Jin, J. Dauwels, S. S. Cash, and M. B. Westover. EEG CLas-

sification via convolutional neural networkbased interictal epileptiform event detection 2018 .The
development of an automated, fast, and reliable epileptic EEG diagnostic system is essential. Interic-
tal epileptiform discharges (IEDs) are recurring patterns that are highly suggestive of epilepsy. The
proposed system achieved a mean 4-fold classification accuracy of 83.86

Nurislamova, Y. M., Novikov, N. A., Zhozhikashvili, N. A., Chernyshev Enhanced thetaband

coherence between midfrontal and posterior parietal areas reflects postfeedback adjustments in the
state of outcome uncertainty 2019, Execution of cognitively demanding tasks requires proper func-
tioning of both task-specific and non-specific neural processes. Performance monitoring system is also
essential for flexible adaptive behavior we investigated functional interaction between performance
monitoring and decision-making systems of the brain. We showed that certain stages of decision mak-
ing process involve the posterior parietal regions. We interpret this as a hallmark of top–down signal
that originates in the midfrontal region.

B.XuL. Zhang, Song, C. Wu, W. Li, D. Zhang, G. Xu, H. Li, and H. Zeng Wavelet transform
timefrequency image and convolutional networkbased motor imagery EEG classification 2019 , Deep

7
learning can be applied to biomedical signals, especially EEG signals. The accuracy on dataset III
from BCIcompetition II reaches 90

P. Zhang, X. Wang, W. Zhang, and J. Chen Learning spatial spectral temporal EEG features
with recurrent 3D convolutional neural networks for cross-task mental workload assessment 2020 Men-
tal workload assessment is essential for maintaining human health and preventing accidents. This 3
paper proposes a structure of deep recurrent and convolutional neural networks to learn EEG features
across different tasks without prior knowledge. The proposed model is validated based on the EEG
signals The original motor imagery electroencephalography (MIEEG) data contains a large number
of spatial features related to the distribution of electrodes on the brain. Most of the current convolu-
tional neural network (CNN) based methods do not take full advantage of the spatial features.

RAMESH BABU VALLABHANENI, PANKAJ SHARMA, VINIT KUMAR Deep Learn-

ing Algorithms in EEG Signal Decoding Application 2021 Deep learning algorithms are becoming
increasingly powerful tool in biomedical engineering. Increasing focus on using them for decoding
physiological or pathological status of the brain from electroencephalographic (EEG) signals. Deep
learning algorithms are increasingly powerful tool in biomedical engineering. Increasing focus on us-
ing them for decoding physiological or pathological status of the brain from electroencephalographic
(EEG) signals. Existing applications of deep learning on EEG include braincomputer interfaces, cog-
nitive neuroscience and diagnosis of brain disorders.

8
 CHAPTER 3

METHODOLOGY

We used a mental workload dataset termed STEW: Simultaneous Task EEG Workload
dataset in this study. This is a publicly available dataset that uses a benchmark test to infer work-
load. The dataset comprises of 50 participants’ pre- and post-workload recordings. The majority of
the brain is represented by these 14 channels. Six channels cover the frontal cortex, two channels
cover the temporal cortex, two channels cover the parietal cortex, two channels cover the occipital
cortex, and two channels cover the motor cortex.The workload assessment pipeline is implemented,
and the baseline for two and three level mental workload assessment is validated. Then we look at
the factors that influence evaluation accuracy. The sliding window size, sliding window step size, and
artefact subspace reconstruction (ASR) threshold will all be examined in this experiment. To estimate
mental workload from each region of the brain, we will conduct a region-based analysis to determine
which brain regions have a larger presence of workload-related information. We will do a frequency
bands analysis to determine which frequency bands include more task-related information in order to
estimate mental workload from each frequency band.

3.1 Data Collection

We used the STEW: Simultaneous Task EEG Workload dataset [34] as a mental workload
dataset in this study. This is a publicly available dataset that uses a benchmark test to infer workload.
The dataset comprises of 50 participants’ pre- and post-workload recordings. The number of people
who took part is large enough. This data was collected using the EMOTIV EPOC+, a 14-channel
EEG instrument. The majority of the brain is represented by these 14 channels. Six channels cover
the frontal cortex, two channels cover the temporal cortex, two channels cover the parietal cortex,
two channels cover the occipital cortex, and two channels cover the motor cortex.In order to acquire
data Lim et al. recruited 50 university graduate students, all of whom had no history of neurolog-
ical or brain-related disorders. In terms of workload, During induction, they performed a SIMKAP
(Simultaneous Capacity) test . This is a GUI-based psychological evaluation test that evaluates mul-
titasking.The test is split into two sections in the GUI. The first is a comparison window, in which

9
subjects must select identical terms or things from one comparison window and mark that term in
the other. The auditory response panel is the second part of the test, in which subjects must choose
an answer to an auditory question. Subjects must complete both tasks at the same time for a total
of 18 minutes.

3.1.1 Data Filtering

In its raw form, the data obtained from EEG recordings is useful. There could be a lot
of noise and unwelcome artefacts in this. It contains different frequencies, but we only require data
between 1 and 50 Hz. Because the essential information is contained within five frequency bands,
the frequency range of these bands is from 1 to 50 Hz, frequencies below 1 Hz and beyond 50 Hz are
worthless. As a result, the data is first high pass filtered at 1 Hz. All frequencies below the threshold
are blocked by the high pass filter, while all frequencies above the threshold are allowed to pass.EEG
signals are prone to noise and must first be cleansed. High amplitude artefacts, electrical line noise,
white noise, and frequencies outside of the necessary ranges are all possible sources of noise in EEG
signals. All of these have an impact on the signal’s useful information. As a result, various types of
filters such as a high pass filter, a low pass filter, a notch/band reject filter, and so on must be applied
to the acquired signal. A high pass filter filters out all frequencies over a predefined threshold, a low
pass filter filters out all frequencies below the threshold, and a notch filter stops a specific frequency
band while allowing all other frequencies to pass through.

3.2 Line noise filtering

Electrical line noise, which happens as a result of electrical equipment operation, has an
impact on signals. As a result, line noise causes misinterpretation in EEG readings, which must be
eliminated. Line noise has a frequency of 60 Hz, hence data for line noise cleaning is 60 Hz.

3.3 Artifact Subspace Reconstruction (ASR)

Artifact subspace reconstruction is a critical challenge for this portion. This is a feature for
removing poor channels and ones with a lot of amplitude and reconstructing them. High amplitude
frequencies might develop as a result of muscular activity or when an electrode is not properly fixed.
The power of signals is increased by these high amplitude aberrations. Because it entirely removes
defective channels, it affects the dimension of the dataset. To clean signals, we only need to remove
high amplitude frequencies.

10
 Figure 3.1: Workload measurement pipeline

3.4 Feature Extraction

So, for the complete dataset, we’ll need some unique identifiers for each signal. Because the
power of the delta, theta, alpha, beta, and gamma frequency bands may be utilised to identify the
existence of workload, these bands are used as distinct aspects of the signals. Signals are first filtered
and split into various bands, and then the power of each signal’s bands is determined. The power
spectral density (PSD) is determined for band power. PSD stands for power spectrum density, which
indicates how much of the total signal power a signal component has.

3.5 Classification

3.5.1 Data labeling

The data must be labelled after the power features have been extracted. As labels, detection
ratings are employed. Despite the fact that the scores range from 0 to 4, we classify this as a three-
class classification problem. This means that ratings ranging from 0 to 1 are considered class one,
indicating a low workload, ratings 2 are considered class 2 indicating a medium workload, and ratings
3-4 are considered class 3 indicating a heavy workload. As a result, the data is divided into three
categories.

11
3.5.2 Training model

Model will be trained on 5-fold cross validation that means 80 percent training data and 20
percent test data for all 5 folds.

3.6 Workload Assessment

3.6.1 Workload detection/recognition

After training of models we will predict the workload. Workload assessment will be done
in all brain regions such as frontal region, parietal region , occipital region individually as well as
combined regions such as frontal and parietal region, frontal and occipital region.Following the work-
load assessment, we will examine the system’s performance in terms of individual region assessment
accuracy, combined region analysis, and computing time.

Figure 3.2: EEG data recording pipelinet

3.6.2 System performance

After workload assessment we will evaluate the system’s performance in terms of assessment
accuracy of individual region as well as combined region analysis and computational time.

12
13
Figure 3.3: Data set-1
14
Figure 3.4: Data set-2
 CHAPTER 4

CODE AND SIMULATION

#IMPORTING LIBRARIES
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.feature_selection import mutual_info_classif
import matplotlib.pyplot as plt
%matplotlib inline
import seaborn as sns

#check your folder

import os
for dirname, _, filenames in os.walk(r"C:\Users\LENIN\UG-EEG"):
for filename in filenames:
print(os.path.join(dirname, filename))

#Importing and Exploring dataset

df=pd.read_csv(r’C:\Users\LENIN\UG-EEG\STEW.csv’)
df

print(df.shape)
print(df.columns)

# These checks help in identifying imbalanced datasets

df[’Detection{0,1}’].value_counts()

15
#Checking the dataset for missing values
for col in df.columns:
if(df[col].isnull().sum()>0):
print(col)
else:
print("No missing value in {}".format(col))

#Checking the total number of Categorical features in dataset

for col in df.columns:
if(df[col].dtype==’object’):
print(col)
else:
print("No Categorical features in dataset {}".format(col))

#Identifying the top-features related to our dataset

#Separating features and columns
df_fea=df.drop(’Detection{0,1}’,axis=1)
y=df[’Detection{0,1}’]

print(df_fea)

print(y)

#mutual_info_classif helps in defining the % of dependence between features and

target variables
mi_score=mutual_info_classif(df_fea,y) #Estimate mutual information for a discrete
target variable.
print(mi_score)

#Converting the labels to [0,1] format

y=pd.get_dummies(df[’Detection{0,1}’])
print(y)

16
#Converting the scores to pandas-series and choosing the columns as the
index of respected score
mi_score=pd.Series(mi_score,index=df_fea.columns)
print(mi_score)

mi_score=(mi_score*100).sort_values(ascending=False)
print(mi_score)

#Selecting features with more than 10% MI-score

#Using the index values(column names of original dataset) to drop the
columns which can interfere with the overall model
top_fea=mi_score[:].index # last 367 columns were dropped because
they were only a liability for our computation power
print(top_fea)

#Scaling the dataset

df_sc=StandardScaler().fit_transform(df_fea[top_fea])
print(df_sc)

#Splitting the dataset into Train ,Test and Validation

Xtr,xte,Ytr,yte=train_test_split(df_sc,y,random_state=108,test_size=0.27)

print(Xtr.shape)
print(Ytr.shape)
print(xte.shape)
print(yte.shape)

print(10935*73/100)
print(10935*27/100)

xtr,xval,ytr,yval=train_test_split(Xtr,Ytr,random_state=108,test_size=0.27)
print(xtr.shape)

17
print(xval.shape)
print(ytr.shape)
print(yval.shape)

#Building/Compiling/Fitting the Neural-Net to identify the emotion

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import callbacks,layers

# Building the model

model=keras.Sequential([
layers.Dense(units=256,input_shape=(Xtr.shape[1],),activation=’relu’),
#using the relu activation because it is great for hidden layers
layers.BatchNormalization(), #BatchNormalization layer scales the
dataset even further
layers.Dropout(0.27), #Dropping-out the nodes to make our model more general
layers.Dense(units=512,activation=’relu’),
layers.BatchNormalization(),
layers.Dropout(0.3),
layers.Dense(units=1024,activation=’relu’),
layers.BatchNormalization(),
layers.Dropout(0.32),
layers.Dense(units=256,activation=’relu’),
layers.BatchNormalization(),
layers.Dropout(0.27),
layers.Dense(units=256,activation=’relu’),
layers.BatchNormalization(),
layers.Dropout(0.32),
layers.Dense(units=256,activation=’relu’),
layers.BatchNormalization(),
layers.Dropout(0.27),
layers.Dense(units=2,activation=’sigmoid’) #Softmax activation
helps in multiclass-identification
])

18
# Compiling the model
adam=keras.optimizers.Adam(learning_rate=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-07)
adamax=keras.optimizers.Adamax(learning_rate=0.00085, beta_1=0.9, beta_2=0.999, epsilon=1e-07)
#Categorical_crossentropy will make sure if all the categories are getting identified
#Accuracy will help in identifying if correct labels are getting picked-up
model.compile(optimizer=adamax,loss=’binary_crossentropy’,metrics=[’accuracy’])

# Fitting the model

call=callbacks.EarlyStopping(patience=10,min_delta=0.0001,restore_best_weights=True)
# Defining earlystopping callback to save time.
history=model.fit(xtr,ytr,validation_data=(xval,yval),batch_size=28,epochs=50,
callbacks=[call])

# summarize history for accuracy

plt.plot(history.history[’accuracy’])
plt.plot(history.history[’val_accuracy’])
plt.title(’Model Accuracy’)
plt.ylabel(’accuracy’)
plt.xlabel(’epoch’)
plt.xlim()
plt.legend([’train’, ’test’], loc=’upper left’)
plt.savefig("Model-Accuracy.jpeg")
plt.show()

# summarize history for loss

plt.plot(history.history[’loss’])
plt.plot(history.history[’val_loss’])
plt.title(’Model loss’,color="green")
plt.ylabel(’loss’)
plt.xlabel(’epoch’)
plt.legend([’train’, ’test’], loc=’lower left’)
plt.savefig("Model-Loss.jpeg")
plt.show()

#Time to Evaluate!

19
model.evaluate(xte,yte, verbose=1)

model.predict(xte)

import numpy as np
prediction=[1 if np.argmax(a, axis=0)>0.5 else 0 for a in model.predict(xte)]
print(prediction)

20
 CHAPTER 5

RESULT AND DISCUSSION

We simulated the workload assessment pipeline for two analyses in this experiment. The
first analysis looked at two levels of workload assessment: class 1 (no or low workload) and class 2
(some or excessive workload). the parameter values that were used. 60 to 79.8 percent classification
accuracy was attained. The impact of several parameters on assessment accuracy was investigated.
The ASR threshold, window size, and overlap size are among them.
we employed ten different ASR threshold values as well as the [512,384] window and overlaps.
size, accordingly, to see how the accuracy of the judgement varies. It can be found in Table 4.2.
The influence of the ASR threshold on accuracy is unpredictable, whereas the maximum accuracy is
constant. 60 to 79.8 percent of the target has been met.

Figure 5.1: Code snippet

21
Figure 5.2: Prediction

22
Figure 5.3: Output

23
Figure 5.4: Model accuracy

24
Figure 5.5: Model loss

25
Figure 5.6: Accuracy

26
Figure 5.7: Features

27
Figure 5.8: Imbalance Dataset

28
 CHAPTER 6

CONCLUSION

In our Proposed Work a mental workload classification method based on EEG features for
vision and operation tasks. Here we used a Deep learning technique which is known as Artificial
Neural Network(ANN). This method, named as Method 1, contains three steps that is filtering the
measured EEG signals , extracting their energy features, and classifying them. The energy features
of EEG are implemented to overcome the problem of low accuracy.The proposed method can realize
fast, accurate, and automatic mental workload classification. Based on the comparison results of our
classification methods, some conclusions can be drawn as follows:Compared with Method 1 based on
EEG signal features, the classification accuracy of our presented method is improved from 60 to 79.8
percent. The classification Loss will be the average difference is 30 percent.

29
 CHAPTER 7

FUTURE ENHANCEMENTS

We have been working on EEG-based workload detection based on an offline dataset in

which workload is produced by a benchmark SIMKAP test in our present research. Using the pro-
posed pipeline, we anticipate new directions for real-time workload detection. The real-time workload
detection pipeline can then be tested for jobs that are performed in real time. the natural surround-
ings This will enable operators’ mental workload to be estimated. in the context of their professional
environment, completing crucial duties Monitoring in real time is more difficult than monitoring in
the past. Because of the intended The pipeline is intended for offline processing and should not be
used to detect real-time workloads.in its current state form. For real-time processing, we need to
record EEG activity for at least 2.5 minutes, which is the length of the STEW dataset recording time.
This 2.5-minute data set will take some time to process and cannot be processed in its entirety. As
a result, we may use parallel processing to speed up data processing. We can also make use of this.
implementation in a real-life setting The sensitivity of real-world scenarios is higher. external noises,
such as muscle and head movement, eye blinks, and so on As a result, in the real world For noise
reduction, monitoring may necessitate a lot more processing.

30
REFERENCES

[1] H. Mei and X. Xu, “EEG-based emotion classification using convolutional neural network,” in
Proc. Int. Conf. Secur., Pattern Anal., Cybern. (SPAC), Dec. 2017, pp. 130–135.

[2] S. Tripathi, “Using deep and convolutional neural networks for accurate emotion classification on
DEAP dataset,” in Proc. 31st AAAI Conf. Artif. Intell., 2017, pp. 4746–4752

[3] Fares Al-Shargie, Tong Boon Tang, and Masashi Kiguchi. Assessment of mental stress effects on
prefrontal cortical activities using canonical correlation analysis: an fNIRS-EEG study. Biomedical
Optics Express, 8(5):2583, 2017. ISSN 2156- 7085. doi: 10.1364/boe.8.002583.

[4] LR Quitadamo, F Cavrini, L Sbernini, F Riillo, L Bianchi, S Seri, and G Saggio. Support vector
machines to detect physiological patterns for eeg and emg-based human–computer interaction: a
review. Journal of neural engineering, 14(1): 011001, 2017.

[5] Hafeez Ullah Amin, Wajid Mumtaz, Ahmad Rauf Subhani, Mohamad Naufal Mohamad Saad, and
Aamir Saeed Malik. Classification of EEG signals based on pattern recognition approach. Frontiers
in Computational Neuroscience, 11 (November):1–12, 2017. ISSN 16625188.

[6] Zhendong Mu, Jianfeng Hu, and Jianliang Min. Driver fatigue detection system using electroen-
cephalography signals based on combined entropy features. Applied Sciences (Switzerland), 7(2),
2017. ISSN 20763417. doi: 10.3390/app7020150.

[7] H. Zeng, C. Yang, G. Dai, F. Qin, J. Zhang, and W. Kong, “EEG classification of driver mental
states by deep learning,” Cogn. Neurodyn., vol. 12, pp. 597–606, Dec. 2018.

[8] S. Chambon, “A deep learning architecture to detect events in EEG signals during sleep,” in Proc.
IEEE 28th Int. Workshop Mach. Learn. Signal Process. (MLSP), Sep. 2018, pp. 1–6.

[9] n Kyoung Choi, Seung Min Lee, Jun Su Ha, and Poong Hyun Seong. Development of an eeg-based
workload measurement method in nuclear power plants. Annals of Nuclear Energy, 111:595–607,
2018.

[10] R. D. Dias, M. C. Ngo-Howard, M. T. Boskovski, M. A. Zenati, and S. J. Yule. Systematic

review of measurement tools to assess surgeons’ intraoperative cognitive workload. British Journal
of Surgery, 105(5):491–501, 2018. ISSN 13652168. doi: 10.1002/bjs.10795.

31
[11] V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M Gordon, C. P. Hung, and B. J. Lance,
“EEGNet: A compact convolutional neural network for EEGbased brain–computer interfaces,” J.
Neural Eng., vol. 15, no. 5, 2018, Art. no. 056013

[12] Sanay Muhammad Umar Saeed, Syed Muhammad Anwar, Muhammad Majid, Muhammad
Awais, and Majdi Alnowami. Selection of Neural Oscillatory Features for Human Stress Clas-
sification with Single Channel EEG Headset. BioMed Research International, 2018, 2018. ISSN
23146141.

[13] Yisi Liu, Zirui Lan, Han Hua Glenn Khoo, King Ho, Holden Li, Olga Sourina, and Wolfgang
Mueller-Wittig. EEG-based evaluation of mental fatigue using machine learning algorithms. Pro-
ceedings - 2018 International Conference on Cyberworlds, CW 2018,

[14] J. Thomas, L. Comoretto, J. Jin, J. Dauwels, S. S. Cash, and M. B. Westover, “EEG CLassifica-
tion via convolutional neural networkbased interictal epileptiform event detection,” in Proc. 40th
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Jul. 2018, pp. 3148–3151.

[15] Rebecca L. Charles and Jim Nixon. Measuring mental workload using physiological measures: A
systematic review. Applied Ergonomics, 74(September 2018):221– 232, 2019. ISSN 18729126. doi:
10.1016/j.apergo.2018.08.028.

[16] Lan. EEG-based emotion recognition using machine learning techniques. Doctoral thesis, Nanyang
Technological University, Singapore, 2018.

[17] ipo Wang Lim, Wei Lun, Olga Sourina. STEW: Simultaneous Task EEG Workload Dataset, 2018.

[18] Nurislamova, Y. M., Novikov, N. A., Zhozhikashvili, N. A., Chernyshev, B. V. (2019). Enhanced
theta-band coherence between midfrontal and posterior parietal areas reflects post-feedback ad-
justments in the state of outcome uncertainty. Frontiers in Integrative Neuroscience, 13(14), 1–14.

[19] B. Xu, L. Zhang, A. Song, C. Wu, W. Li, D. Zhang, G. Xu, H. Li, and H. Zeng, “Wavelet transform
time-frequency image and convolutional network-based motor imagery EEG classification,” IEEE
Access, vol. 7, pp. 6084–6093, 2019

[20] P. Zhang, X. Wang, W. Zhang, and J. Chen, “Learning spatial-spectraltemporal EEG features
with recurrent 3D convolutional neural networks for cross-task mental workload assessment,” IEEE
Trans. Neural Syst. Rehabil. Eng., vol. 27, no. 1, pp. 31–42, Jan. 2019.

[21] Sami Elzeiny and Marwa Qaraqe. Machine Learning Approaches to Automatic Stress Detection:
A Review. Proceedings of IEEE/ACS International Conference on Computer Systems and Appli-
cations, AICCSA, 2018-Novem:1–6, 2019. ISSN 21615330. doi: 10.1109/AICCSA.2018.8612825

[22] Mike X Cohen. Analyzing Neural Time Series Data. Analyzing Neural Time Series Data, 2019.

32
[23] Sanay Muhammad Umar Saeed, Syed Muhammad Anwar, Humaira Khalid, Muhammad Majid,
and Ulas Bagci. Electroencephalography based Classification of Long-term Stress using Psycho-
logical Labeling. Sensors 20, no. 7: 1886, pages 1–8, 2019

[24] Ethem Alpaydin. Introduction to machine learning. MIT press, 2020.

33